1 /**
2 * Copyright 2023 Huawei Technologies Co., Ltd
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include "nnacl/nnacl_kernel.h"
18 #include "nnacl/cxx_utils.h"
19 #include "src/tensor.h"
20 #include "include/errorcode.h"
21 #include "nnacl/errorcode.h"
22
23 using mindspore::lite::RET_ERROR;
24 using mindspore::lite::RET_OK;
25 namespace mindspore::nnacl {
~NNACLKernel()26 NNACLKernel::~NNACLKernel() {
27 if (in_ != nullptr) {
28 free(in_);
29 in_ = nullptr;
30 }
31 if (out_ != nullptr) {
32 free(out_);
33 out_ = nullptr;
34 }
35
36 if (kernel_ != nullptr) {
37 kernel_->Release(kernel_);
38
39 free(kernel_);
40 kernel_ = nullptr;
41 }
42 }
43
Prepare()44 int NNACLKernel::Prepare() {
45 if (kernel_ == nullptr) {
46 return RET_ERROR;
47 }
48
49 int ret = kernel_->Prepare(kernel_);
50 if (ret != RET_OK) {
51 MS_LOG(ERROR) << "NNACL prepare failed. Kernel: " << name() << ", ret: " << ret;
52 MS_LOG(ERROR) << NNACLErrorMsg(ret);
53 return ret;
54 }
55
56 if (!InferShapeDone()) {
57 return RET_OK;
58 }
59 return ReSize();
60 }
61
ReSize()62 int NNACLKernel::ReSize() {
63 if (kernel_ == nullptr) {
64 return RET_ERROR;
65 }
66 UpdateTensorC();
67
68 int ret = kernel_->Resize(kernel_);
69 if (ret != RET_OK) {
70 MS_LOG(ERROR) << "NNACL resize failed. Kernel: " << name() << ", ret: " << ret;
71 MS_LOG(ERROR) << NNACLErrorMsg(ret);
72 return ret;
73 }
74 return RET_OK;
75 }
76
Run()77 int NNACLKernel::Run() {
78 if (kernel_ == nullptr) {
79 return RET_ERROR;
80 }
81 UpdateTensorC();
82 kernel_->workspace_ = workspace();
83
84 int ret = kernel_->Compute(kernel_);
85 if (ret != RET_OK) {
86 MS_LOG(WARNING) << "NNACL compute failed. Kernel: " << name() << ", ret: " << ret;
87 MS_LOG(WARNING) << NNACLErrorMsg(ret);
88 return ret;
89 }
90 return RET_OK;
91 }
92
UpdateTensorC()93 void NNACLKernel::UpdateTensorC() {
94 for (size_t i = 0; i < in_size_; i++) {
95 in_[i] = in_tensors().at(i)->ConvertToTensorC();
96 }
97 for (size_t i = 0; i < out_size_; i++) {
98 out_[i] = out_tensors().at(i)->ConvertToTensorC();
99 }
100 }
101
OptimizeDataCopy()102 int NNACLKernel::OptimizeDataCopy() {
103 auto input_tensor = in_tensors().front();
104 CHECK_NULL_RETURN(input_tensor);
105 CHECK_NULL_RETURN(input_tensor->data());
106 auto output_tensor = out_tensors().front();
107 CHECK_NULL_RETURN(output_tensor);
108 CHECK_NULL_RETURN(output_tensor->data());
109
110 if (input_tensor->allocator() == nullptr || input_tensor->allocator() != output_tensor->allocator() ||
111 input_tensor->allocator() != ms_context_->allocator || /* runtime allocator */
112 op_parameter_->is_train_session_ || !output_tensor->own_data()) {
113 return NNACLKernel::Run();
114 }
115
116 output_tensor->FreeData();
117 output_tensor->ResetRefCount();
118 output_tensor->set_data(input_tensor->data());
119 if (input_tensor->IsConst()) {
120 output_tensor->set_own_data(false);
121 } else {
122 output_tensor->set_own_data(input_tensor->own_data());
123 }
124 return RET_OK;
125 }
126
NNACLCheckArgs()127 int NNACLKernel::NNACLCheckArgs() {
128 if (op_parameter_ == nullptr) {
129 MS_LOG(ERROR) << "NNACL check failed. Invalid parameter.";
130 return RET_ERROR;
131 }
132
133 if (in_size_ == 0 || in_size_ * sizeof(TensorC *) > MAX_MALLOC_SIZE) {
134 MS_LOG(ERROR) << "NNACL check failed. Invalid input size: " << in_size_;
135 return RET_ERROR;
136 }
137
138 if (out_size_ == 0 || out_size_ * sizeof(TensorC *) > MAX_MALLOC_SIZE) {
139 MS_LOG(ERROR) << "NNACL check failed. Invalid output size: " << out_size_;
140 return RET_ERROR;
141 }
142
143 if (op_parameter_->thread_num_ <= 0 || op_parameter_->thread_num_ > MAX_THREAD_NUM) {
144 MS_LOG(ERROR) << "NNACL check failed. Invalid thread number: " << op_parameter_->thread_num_;
145 return RET_ERROR;
146 }
147
148 return RET_OK;
149 }
150
InitKernel(const TypeId & data_type,const lite::InnerContext * ctx)151 int NNACLKernel::InitKernel(const TypeId &data_type, const lite::InnerContext *ctx) {
152 CHECK_NULL_RETURN(ctx);
153
154 in_size_ = in_tensors_.size();
155 out_size_ = out_tensors_.size();
156
157 int ret = NNACLCheckArgs();
158 if (ret != RET_OK) {
159 MS_LOG(ERROR) << "NNACL check args failed. Kernel: " << name();
160 return ret;
161 }
162
163 in_ = reinterpret_cast<TensorC **>(malloc(in_size_ * sizeof(TensorC *)));
164 if (in_ == nullptr) {
165 return RET_ERROR;
166 }
167
168 out_ = reinterpret_cast<TensorC **>(malloc(out_size_ * sizeof(TensorC *)));
169 if (out_ == nullptr) {
170 return RET_ERROR;
171 }
172
173 UpdateTensorC();
174 kernel_ = CreateKernel(op_parameter_, in_, in_size_, out_, out_size_, data_type, exec_env_);
175 if (kernel_ == nullptr) {
176 MS_LOG(WARNING) << "NNACL create kernel failed. Kernel: " << name();
177 return RET_ERROR;
178 }
179 kernel_->UpdateThread = DefaultUpdateThreadNumPass;
180 return RET_OK;
181 }
182 } // namespace mindspore::nnacl
183